Cloud Hosting

Side stability of open pit mine

When developing open-pit mines, the limitation of rock slope safety is the main reason that affects mine production efficiency. The proportion of open-pit mining in China is actually very large. The mining of iron ore and fossil raw materials is almost always in the form of open pit mining. During the mining process, the safety of the slope body is the most important. Therefore, in the mining process, it is necessary to increase the final slope angle and ensure the stability of the slope. There will be a very sharp contradiction in the mining process, that is, the larger the final slope angle, the more unstable the slope will be. If this problem is not handled properly, it will seriously affect the safety production of the mine and the economic benefits of the mine. And the mining of open-pit mines is very likely to cause the surrounding environment to become unsafe. Therefore, when mining, it is necessary to ensure safety without reducing the mining speed and reducing the impact on the surrounding environment. Committed to achieving an economical and efficient stripping ratio10. The mining of the open pit mine is shown in Fig.1.

Mining of an open pit mine.

As shown in Fig.1, in the mining process of the mine, professional tools such as excavators need to be used11. And it will form a first-order slope. In fact, open-pit mines in China are still relatively common, and there are many mines in Inner Mongolia and Xinjiang. These mines have very serious stability problems, and many large landslides have occurred. This has caused a lot of economic losses to the mining mines. So the issue of stability has always been an essential issue. At the same time, because the shallow resources are being developed by people, they are constantly decreasing. Therefore, new technologies have been created to develop deep resources. However, this technology will have a very large impact on the entire mine rock formation, the environment will also be damaged, and safety and stability will also be reduced. Although many scholars have improved the stability of mining in different ways, the stability of each mine is an independent quantity, and they have different characteristics. Therefore, it is necessary to determine which method to use for prediction according to the on-site assessment of the mine.Therefore, further research is needed to find better ways to mine mines12.

If the stability of the slope is not good, it is easy to occur a landslide disaster, which is a very serious problem13. It threatens people's property and lives. The occurrence of landslides is generally due to the destruction of rock slopes. Rockslides and rockslides are the main types of rock slope damage as shown in Picture 2:

As shown in Fig.2, there are two main types of landslides14. The first is rock avalanches. It mostly happens on the kind of very steep slopes where the rock breaks apart in chunks and then collapses, tumbling forward. The rock body at the top is often detached and then falls off due to some factors, and accumulates at the foot of the slope. These situations often occur where there are cracks on the top of the slope. Cracks are also created by weathering of rock over time, or by the intrusion of rainwater and prolonged soaking. However, it is also possible that due to changes in temperature, high temperature or shading may cause the rock to loosen. The protective measures taken by general experts are to use artificially reinforced building materials, that is, anchor cables. This way, the impact force of rock mass collapse and sliding can be minimized. The second is rock slip, which is a phenomenon in which the rock mass slides along a certain surface15. In fact, the main reason for rock slip is because of too much rainfall. After surface water seeps into the cracks, it will generate hydrostatic pressure, which is the force that promotes the sliding of the soil slope and is detrimental to the stability of the soil slope. Due to the infiltration of rainwater, the rise of river water level, or the impoundment of reservoirs, the groundwater level rises, causing static water pressure to act on the impermeable structural surface of the slope. It acts perpendicular to the structural surface and acts on the slope, weakening the normal stress generated by the weight of the sliding mass on the surface, thereby reducing the antisliding resistance of the soil. There are several types of rock slides, so I wont introduce them one by one here. Generally speaking, rock sliding is plane sliding. It means that when the rock slides along the plane, the plane is more prone to plane sliding when the inclination angle of the sliding surface is greater than the internal friction angle. Two conditions need to be satisfied for the plane sliding of the slope rock mass, that is, to overcome the resistance on both sides and the resistance at the bottom. In soft rock, when the bottom inclination angle of the slope rock mass is much larger than the internal friction angle of the rock mass in the open-pit mine slope rock mass, the lateral restraint of the rock mass cannot provide enough force to prevent the rock from being damaged. Will detach from the slope rock mass to produce plane sliding. In the hard rock slope rock mass, only when the discontinuous surface of the slope rock mass crosses the top of the slope, and the rock on the slope is separated from the rock on both sides, the slope rock mass without lateral restraint may also slide in a plane16.

Rock mass characteristics are another tool for classifying slopes, especially in mines. SMR is the most common classification scheme and is often used by different researchers to analyze the stability of cutting slopes in different mines. Slope quality rating is the main tool for understanding the rock mass behavior of open-pit mine slopes. Due to the increase in depth and slope angle, slope quality rating always brings serious problems. Due to various geological complexities, stability issues are more severe. The stability analysis of the moving slope was conducted using the Stereonet diagram. It is a simple tool to analyze wedge failure in planar and rock slopes. This structural data is geometrically plotted to establish the failure probability of the equal area network in the pattern17. There are also many ways to control the slope, there are generally three methods. As shown in Fig.3.

Methods of treating slopes.

From Fig.3, it can be clearly seen that these three methods of slope management18. The first method is to dig up and make up. The general meaning is that there will be many rock masses with poor stability near the upper part of the slope. These rock masses with poor stability can be dug up, transported to the foot of the slope, and compacted. This can effectively enhance the stability. However, because the traction between the rock masses is still very strong, only the rock masses with poor stability can be dug up. The second method is drainage. Because rain is an essential reason for affecting stability. The accumulation of rainwater will affect the slippage of the cracks on the rock surface, resulting in the occurrence of landslides. Especially in the treatment of high, steep, and large slopes, drainage is particularly important. The third method is to use artificial structures for reinforcement. Anchor cables are generally used for protection and reinforcement. Of course, there are also retaining walls and antislide piles. All three methods work well. When controlling slopes, they can be used in combination to achieve better results19.

Moreover, slope material is important or slope geometry is important. Classified by stratigraphic lithology: it can be divided into soil slopes and rock slopes. (a) According to the rock structure, it is divided into layered structure slope, block structure slope, and network structure slope; (b) According to the relationship between rock strata inclination and slope direction, it can be divided into forward slope, reverse slope, and vertical slope. All slope instability involves the failure of slope rock and soil under shear stress. Therefore, the factors that affect the shear stress and the shear strength of rock and soil all affect the stability of the slope.

Deep learning is actually a kind of machine learning method, and its predecessor is machine learning and artificial neural network20. However, because of the passage of time, this method is constantly developing and optimizing, and its application fields are also very wide. Specifically as shown in Fig.4.

Application areas of deep learning.

As shown in Fig.4, this method has many application fields. The author lists nine areas in total. First of all, in the field of computer vision, this method can help computers process image data, or recognize text, and convert these images or text, which is very intelligent and convenient21. In the field of speech recognition, with the support of this algorithm, the efficiency of speech recognition has been greatly improved. Just like processing image data in computer vision, the method can turn sounds into recognizable models very quickly. In the field of audio recognition, this method is also used to improve the efficiency of audio recognition. In terms of social network filtering, the components in the network are very messy, and there are all kinds of information, but this method is very good at information classification, so this method is also very suitable for filtering social networks. In terms of machine translation, using this method can improve the quality of machine translation and make machine translation more inclined to the translation level of an ordinary person. In drug design, this method can assist the development of small molecule drugs, provide new computational decisions for pharmaceuticals, and process more chemical data information. In bioinformatics, using this method can bring new changes to the discipline. Because the method is so good at mining data, it is well suited for solving genomics problems. In the field of medical image analysis, after applying this method, a fast and very detailed analysis of medical images can be performed better. Because this method has already achieved good results in image segmentation. Therefore, it is also very suitable for image analysis in the field of medical images22.

The deep learning method not only has a wide range of applications, but also has many advantages, as shown in Fig.5.

Advantages of deep learning.

As shown in Fig.5, its first advantage is high versatility. Generally, the data we deal with are multidimensional ordered data, and then due to the rapid development of big data, deep learning has been applied very well in various fields23. In addition to speech recognition and image classification, it also has very good performance in data mining and data processing and data prediction, so its versatility is high. The second advantage is robustness, which means that the method is smarter and more stable24. It can automatically adjust parameters according to data changes and automatically adapt to data changes. The third advantage is a good generalization. After the data is increased, it can still have good generalization ability, and the performance is not weakened at all, but enhanced. The fourth advantage is scalability, because when the neural network is stacked too much, the gradient will disappear or the gradient will explode, and this method can solve this problem very well. And this method has very good scalability in the number of layers and structural parameters and can be freely combined to achieve a better learning effect25.

In addition, the method can be generalized to the neural network structure trained in different fields, and can also have a good training effect in the case of insufficient data. We can compare the performance of machine learning and deep learning at training time, as shown in Fig.6.

The relationship between the amount of training data and training performance.

As shown in Fig.6, it is obvious that the previous machine learning has too few parameters, and when the training data increases, the generalization ability will decrease26. The method proposed in this paper, when the training data increases, the generalization ability is better. This shows that the method proposed in this paper not only has good stability in data processing, but also can process very well data at the same time. Not affected by the size of the data volume.

Since deep learning is widely used, its framework has also been introduced by scholars. The code of the framework itself is very concise, the supported language types are quite rich, the technical documentation is complete, and the maintenance and operation are in good condition. Below I will list five more popular frameworks, as shown in Table 1.

As shown in Table 1, there are actually many mainstream frameworks, but for the convenience of analysis, the author lists the five most popular frameworks. The first framework is Tensor Flow, and its core code is written in C++. This is generally used to deal with multidimensional vectors. This framework also has visualization tools that can fully display the structure and data flow of the neural network. It contains many mainstream algorithms, and the entire design process is also comprehensive, which is very suitable for prediction in the industry. The second framework is Caffe, which defines each neural network. After ensuring normal docking, the network construction work is just stacking each layer. And it can participate in training as long as the model is defined, and the training performance is very good. The third framework is Torch, whose popularity is mainly due to the support of Facebook. This framework supports a lot of scientific computing, and it is generally the first choice for scientific research in academia. The fourth is CNTK, which is introduced by Microsoft. There are also many features, mainly the network structure is very fine, the code is product-level, and can be trained on a variety of hardware. The fifth framework is Keras. Its components are highly encapsulated. It is generally used by beginners, and it is relatively quick to get started. After understanding the principle, you can initially build a network.

(1) Recurrent neural network (rnn).

It is mainly a neural network generated to process sequence data.

Assuming the time is Y, you can get the model output as:

$${P}_{Y}=Bcdot {J}_{Y}+{N}_{P}$$

(1)

When the model predicts the output value at time Y, we analyze the loss function, and the backpropagation starts from the final loss value. Then during backpropagation, R represents the cost function, and the defined objective function is:

$$R=Vsum_{Y=1}^{Y}{Vert {A}_{Y}-{U}_{Y}Vert }^{2}=V{sum }_{Y=1}^{Y}sum_{K=1}^{A}{({A}_{Y}left(Kright)-{U}_{Y}left(Kright))}^{2}$$

(2)

The weight formula E can be updated by adjusting the cost function to be smaller:

$${E}^{NEW}=E-rho frac{vartheta R}{vartheta E}$$

(3)

represents the learning efficiency, which can control the speed of parameter update. If it is not properly controlled, it will cause the optimization speed to not keep up. So to calculate the gradient, the error can be formulated as:

$${varepsilon }_{Y}^{U}left(Kright)=-frac{vartheta R}{vartheta {B}_{Y}left(Kright)}$$

(4)

$${varepsilon }_{Y}^{J}left(Kright)=-frac{vartheta R}{vartheta {I}_{Y}left(Kright)}$$

(5)

Then recursively calculate them, and the new formula can be obtained as:

$${U}_{Y}=Hleft({E}_{JU}Gleft({E}_{CJ}{C}_{Y+}{E}_{JJ}{J}_{Y-1}right)right)$$

(6)

$$R=V{sum }_{Y=1}^{Y}{Vert {A}_{Y}-{U}_{Y}Vert }^{2}=Vsum_{Y=1}^{Y}{left({A}_{Y}left(Kright)-{U}_{Y}left(Kright)right)}^{2}$$

(7)

Y represents the last time point, at which the hidden layer can be expressed as:

$${varepsilon }_{Y}^{J}left(Kright)=-left(sum_{O=1}^{A}frac{vartheta R}{vartheta {B}_{Y}(O)}frac{vartheta {B}_{T}(0)}{vartheta {J}_{Y}(O)}frac{vartheta {J}_{Y}(K)}{vartheta {I}_{Y}(K)}right)$$

(8)

By derivation of this formula, the error formula at other time points can be obtained as:

$${varepsilon }_{Y}^{J}left(Kright)=left({O}_{Y}left(Kright)-{U}_{Y}left(Kright){H}{prime}left({B}_{Y}left(Kright)right)right)$$

(9)

The error formulas in the output layer and hidden layer are:

$${varepsilon }_{Y}^{J}left(Kright)=left[sum_{O=1}^{M}{varepsilon }_{Y+1}^{J}left(Oright){E}_{HH}left(O,Kright)+sum_{O=1}^{L}{varepsilon }_{Y}^{U}left(Oright){E}_{JU}left(O,Uright)right]{G}{prime}left({I}_{Y}left(Kright)right)$$

(10)

$${varepsilon }_{Y}^{J}=left[{E}_{JJ}^{Y}{varepsilon }_{Y+1}^{J}+{E}_{JU}^{Y}{varepsilon }_{Y}^{U}right]cdot {G}{prime}left({I}_{Y}right)$$

(11)

It represents the error at time point Y and represents the error at time Y+1.

This way, the weights of the output layer can be updated as:

$${mathrm{E}}_{mathrm{JU}}^{mathrm{NEW}}left(mathrm{O},mathrm{K}right)={mathrm{E}}_{mathrm{JU}}left(mathrm{O},mathrm{K}right)-upbeta {sum }_{mathrm{Y}=1}^{mathrm{Y}}{upvarepsilon }_{mathrm{Y}}^{mathrm{U}}left(mathrm{O}right){mathrm{J}}_{mathrm{Y}}left(mathrm{K}right)$$

(12)

The weights of the input layer can be updated as:

$${E}_{CU}^{NEW}left(O,Kright)={E}_{CU}left(O,Kright)-beta {sum }_{Y=1}^{Y}{varepsilon }_{Y}^{J}left(Oright){C}_{Y}left(Kright)$$

(13)

The weights of the recurrent layer can be updated as:

$${E}_{JJ}^{NEW}left(O,Kright)={E}_{JJ}left(O,Kright)-beta {sum }_{Y=1}^{Y}{varepsilon }_{Y}^{J}left(Oright){J}_{Y-1}left(Kright)$$

(14)

(2) Long short-term memory neural network (lstm)

Although it is similar in structure to RNN, it will change the increased cell state in the structure according to the existence time, and this cell state is a long-term memory27. The LSTM algorithm is often used to perform operations such as prediction of various data or image recognition. Its forgetting gate determines whether the knowledge I have already learned is useful, and which part I want to discard; the input gate determines whether the knowledge others tell me is useful to me, and which knowledge I want to receive; integrating my current knowledge through the output gate some knowledge determines what to report to others28. The theory and learning process of LSYM are valuable, and it has good problem-solving ability when solving some practical problems.

If it is a parameter in the output gate, it is the output of the hidden layer. The formula of the input gate can be obtained as:

$${O}_{Y}=tau left({E}_{CO}{C}_{Y}+{E}_{DO}{D}_{Y-1}+{E}_{VO}{V}_{Y-1}+{N}_{P}right)$$

(15)

$${D}_{Y}={P}_{Y}TANH({V}_{Y})$$

(16)

At this time, if you want to update the cell state, the formula can be expressed as:

$${V}_{Y}={G}_{Y}{V}_{Y-1}+{O}_{Y}{V}_{Y}$$

(17)

$$overline{{V }_{Y}}=tau left({E}_{O}cdot left[{J}_{Y-1},{C}_{1}right]+{N}_{O}right)$$

(18)

Calculated according to the state of the current time, the final output value can be obtained.

$${P}_{Y}=tau left({E}_{CP}{C}_{Y}+{E}_{DP}{J}_{D-1}+{N}_{P}right)$$

(19)

$${D}_{Y}={P}_{Y}TANH({V}_{Y})$$

(20)

Continued here:

Prediction of stability coefficient of open-pit mine slope based on ... - Nature.com